Traffic Congestion Detection and Optimizing Traffic Flow

Using Object Detection, Optical Flow and Fluid Dynamics

by

Md. Shahriyar Hossain

21141017

Md. Imtiaz Bhuiyan

18101688

Marjahan Akther Dulali

17301010

A thesis submitted to the Department of Computer Science and Engineering

in partial fulfillment of the requirements for the degree of
B.Sc. in Computer Science

Department of Computer Science and Engineering

Brac University
January 2022

© 2022. Brac University

All rights reserved.
Declaration
It is hereby declared that

1. The thesis submitted is our own original work while completing degree at Brac
University.

2. The thesis does not contain material previously published or written by a

third party, except where this is appropriately cited through full and accurate
referencing.

3. The thesis does not contain material which has been accepted, or submitted,
for any other degree or diploma at a university or other institution.

4. We have acknowledged all main sources of help.

Student’s Full Name & Signature:

Md. Shahriyar Hossain Md. Imtiaz Bhuiyan

21141017 18101688

Marjahan Akther Dulali

17301010

i
Approval
The thesis titled “Traffic Congestion Detection and Optimizing Traffic Flow Using
Object Detection, Optical Flow and Fluid Dynamics” submitted by

1. Md. Shahriyar Hossain (21141017)

2. Md. Imtiaz Bhuiyan (18101688)

3. Marjahan Akhter Dulali (17301010)

Of Fall, 2021 has been accepted as satisfactory in partial fulfillment of the require-
ment for the degree of B.Sc. in Computer Science and Engineering on January 20,
2022.

Examining Committee:

Supervisor: (Member)

Md. Ashraful Alam, PhD

Assistant Professor
Department of Computer Science and Engineering
BRAC University

Program Coordinator: (Member)

Md. Golam Rabiul Alam, PhD

Associate Professor
Department of Computer Science and Engineering
Brac University

Head of Department: (Chair)

Sadia Hamid Kazi, PhD

Chairperson and Associate Professor
Department of Computer Science and Engineering
Brac University

ii
Abstract
Bangladesh has been suffering a severe traffic congestion issue ever since it has been
on a high paced development roadmap. Researches regarding solving such traffic
issue has been in the talks but has never reached a proper conclusion and far from
implementation. It has slowly grown into a towering challenge to overcome. And
with an aim to topple that tower, we propose a 3 layer architecture to solve this
problem. The proposed model consists of object detection, speed measurement and
decision based on traffic flow. Using neural network object detection algorithms,
it will detect congestion and the speed of the congestion. Then, it will use fluid
dynamics based model to get the traffic flow, pass data between other signals and
provide correct traffic signals. All signals would interact with each other like hive
mind to maximize the traffic flow in any intersection. With the working model we
had at our hand, we ran rigorous experiments to check whether our model works or
not. Our results indicate that our model surpasses all other similarly implemented
models by a noticeably large margin.

Keywords: AITS, ITSC, Deep Learning, Image Recognition, Self-Adaptive Traffic

System, ATS, City Traffic, Fluid Dynamics, Numerical Simulation, Traffic Simula-
tion, Object detection, Optical Flow, Traffic Flow

iii
Dedication
Dedicated to the people who work hard for the people of this country.

iv
Acknowledgement
Firstly, all praises to the almighty Allah, by whose will we have completed this thesis
without having to face any major issues or interruptions.
Secondly, to our supervisor Dr. Ashraful Alam for guiding us throughout the thesis
and providing us directions and resources, with the help of which, we were finally
able to complete the thesis properly.
And finally to our parents who have supported us right from the beginning and
showed their kindness during our tough times.

v
Table of Contents

Declaration i

Approval ii

Abstract iii

Dedication iv

Acknowledgment v

Table of Contents vi

List of Figures viii

List of Tables ix

Nomenclature x

1 Introduction 1
1.1 Research Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.1.2 Research Objectives . . . . . . . . . . . . . . . . . . . . . . . 4
1.2 Thesis Orientation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

2 Literature Review 6
2.1 Intelligent Traffic System Control (ITSC) . . . . . . . . . . . . . . . . 6
2.2 Used Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.3 Related Works . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8

3 Methodology 9
3.1 Data Collection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
3.1.1 Object Detection Data . . . . . . . . . . . . . . . . . . . . . . 11
3.1.2 Traffic Simulation Data . . . . . . . . . . . . . . . . . . . . . . 11
3.2 Data Preprocessing . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

vi
 3.3 Model Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.3.1 Global Speed Detection Algorithm . . . . . . . . . . . . . . . 15
3.4 Fluid Dynamics Approach . . . . . . . . . . . . . . . . . . . . . . . . 16
3.4.1 Mathematical Formulation . . . . . . . . . . . . . . . . . . . . 16
3.4.2 Computational Functions . . . . . . . . . . . . . . . . . . . . . 17

4 Implementation and Results 18

4.1 Implementation Roadmap of Proposed Models . . . . . . . . . . . . . 18
4.2 Object Detection Layer . . . . . . . . . . . . . . . . . . . . . . . . . . 19
4.2.1 Training The Object Detection Model . . . . . . . . . . . . . . 19
4.2.2 Inferencing and Results . . . . . . . . . . . . . . . . . . . . . . 22
4.3 Optical Flow Layer . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
4.4 Numerical Simulations Layer . . . . . . . . . . . . . . . . . . . . . . . 24
4.4.1 Test Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
4.5 Traffic Light Simulation . . . . . . . . . . . . . . . . . . . . . . . . . 25

5 Conclusion 28
5.1 Future Works . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28

vii
List of Figures

3.1 Workflow diagram of the proposed model . . . . . . . . . . . . . . . . 10

3.2 Preprocessing Steps Example . . . . . . . . . . . . . . . . . . . . . . 13
3.3 Yolov5 Inferencing Steps . . . . . . . . . . . . . . . . . . . . . . . . . 14

4.1 Training Precision . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20

4.2 Training Recall . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
4.3 Validation Loss . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
4.4 Mean Average Precision . . . . . . . . . . . . . . . . . . . . . . . . . 21
4.5 Inferencing Results of Trained Model . . . . . . . . . . . . . . . . . . 22
4.6 Inferencing Results of Trained Model . . . . . . . . . . . . . . . . . . 23
4.7 Testing Final Volume Method Against Exact Solution . . . . . . . . . 25
4.8 Testing Final Volume Method Against Exact Solution . . . . . . . . . 26

viii
List of Tables

3.1 Dataset Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

3.2 Collected Data For Inferencing . . . . . . . . . . . . . . . . . . . . . . 12
3.3 Dataset Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.4 Riemann Solution Mapping . . . . . . . . . . . . . . . . . . . . . . . 17

4.1 Training Set Results . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

4.2 Validation Set Results . . . . . . . . . . . . . . . . . . . . . . . . . . 20
4.3 Optical Flow Accuracy . . . . . . . . . . . . . . . . . . . . . . . . . . 23
4.4 L2 norm values . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
4.5 Simulation Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . 26

ix
Nomenclature

The next list describes several symbols & abbreviation that will be later used within
the body of the document

AI Artificial Intelligence

BN Bayesian Network

CCT V Closed Circuit Television

CN N Convolutional Neural Network

GP U Graphical Processing Unit

IoT Internet Of Things

IT S Intelligent Traffic System

IT SC Intelligent Traffic System Control

KN N K Nearest Neighbour

LST M Long Short Term Memory

LW R Lighthill-Whiteham-Richards Model

mAP Mean Average Precision

RCN N Recurrent Neural Network

RF ID Radio Frequency Identification

RGB Red Greed Blue

ROI Region Of Interest

Y OLO You Only Look Once

x
Chapter 1

Introduction

Traffic congestion is one of the biggest nightmares for any megacity. Our Dhaka city
is no exception to this as it still has manually regulated traffic systems because of
unplanned road design, weak infrastructure and also for reckless drivers who have
no regard for traffic regulations [15]. According to available statistics, Dhaka in-
habitants spend an average of 170 hours per year on the road, [19] resulting in a
staggering loss of BDT 1 million each day [13]. For this sole reason, modernizing
Bangladesh’s traffic management system has become an urgent need.

There have been numerous studies conducted on technologies such as IoT-based

cloud solutions.[14], [16], [19]. RFID-based vehicle detection, as well as V2I road-
side infrastructure that combines AI and big data. Considering Bangladesh has
many cars that are not equipped with contemporary technology, almost all of these
vehicles missing the smart characteristics that we want to establish an IoT-based
architecture. Which makes implementation of the highly effective, efficient but so-
phisticated much more difficult.[3], [21], [33], [37], [41], [43]

Thankfully, some effective research on forecasting short-term and long-term traf-

fic flow fluctuations using an image-based technique have been published.[34], [39].
Furthermore, anticipating traffic at least 5 to 10 signal cycles beforehand provides
the system with a solid perspective [35] of how the traffic should be managed and
allows the system to proactively modify the traffic signals. In addition, the system
may adjust the length of traffic signals based on traffic load.[5], [36]

This paper will go into detail on the approaches used, such as parametric and non-
parametric models. [36]. Non-parametric models can also be considered as a good
alternative options as recent strides in computational capabilities and data quan-
tity. We can now do several researches with not only Long Short Term Memory,
Recurrent Neural Networks, Convolutional Neural Network, Graph CNN but also

1
faster computation with SVM, KNN, Bayesian Network Random forest [6], [7], [12],
[28], [33], [35], [40], [47], [48].

1.1 Research Problem

Traffic congestion causes huge mental pressure on people because it causes rush
among people[28]. Not to mention the amount of carbon emission occurring from
the vehicles since only a scarce amount of cars are hydrogen cell cars. If it is looked
at statistically, Bangladesh loses an unbelievable amount of resources due to the
traffic problem. Around 37 thousand Bangladeshi taka is lost and approximately 4
million profession working hours are being spent on the road[41]. So it can be implied
that, traffic congestions significantly damages our economy, health and environment.

In an approach to solve the problems, researchers went to various domains of sci-

entific discoveries to find which solutions could be implemented appropriately. One
such domain is the new emerging domain of machine learning and artificial intelli-
gence. They have noticed the possibility of integrating these technologies into these
issues, providing a new way to solve these scenarios [20], [21]. Additionally, a lot
more efforts and investments are already being made into this field in order to make
great strides in solving traffic congestion problems.

To put in simple terms, the traffic system that Bangladesh currently maintains
doesn’t deem to even be acceptable. The system is broken and unusable. We respect
all the traffic policemen for their contribution, but it simply isn’t in any shape to be
even recovered. It has a very heavy latency due to manual human intervention. Not
to mention the recurring miscommunication, and it is not even the start [15]. They
communicate with their counterparts on each intersection, but the overwhelming
traffic flow becomes infeasible for human interpretation [30]. In addition to that,
fatigue catches up to them quite easily due to rough environment and weather
conditions, and they lose motivation to conduct the traffic flow properly [28]. But if
we can imagine a big centralized AI controlling everything with end to end control,
we can easily avoid all the drawbacks we mentioned above. Of course, it is not just
traffic police that we are considering weakness. The city has heavy concentration
of vehicle, no strict rules, malfunctioning equipments are just to name a few [33].
Either way, we can safely assess that machine based solutions would perform the
greatest optimization for a scenario like this.

2
1.1.1 Motivation
To the following issue, why can’t we adopt IoT-based alternatives? [11], [14], [27],
[29]? Bangladesh has about 3.4 million registered and 3 lacs unregistered cars, ac-
cording to the BRTA database.Almost all of these vehicles lack the necessary modern
technologies for an IoT-based solutions.[10], [37]. However, it is not viable to initi-
ate the process of modernizing automobiles by attaching IoT devices to them right
now. It would not only be an extraordinarily expensive infrastructure, but it will
also be highly unlikely given the country’s current status. [14]. As a result, an
immediate and efficient substitute approach is necessary. With all that in mind,
the question like which appropriate option to select for the best solution emerges.
Is this even possible to find remedies for Dhaka? Yes, considerable study has been
undertaken on this issue in order to select the most appropriate image-based so-
lution customized to the demands of each unique city in order to solve their own
set of traffic challenges.[4], [15], [21], [27], [30], [34], [39], Not only have academics
discovered several methods for analyzing urban flow of traffic, but some methods
have also been substantially deployed.[1], [2], [16].

A deep learning-based traffic state categorization approach based on our studied

articles is suggested to identify congestion problems at urban crossings. A popular
method is to depend on numerical data such as vehicle flow [40], [47]and speed.[40],
[47]. Furthermore, utilizing the YOLOv3,CNN, R-CNN,HOG, or SSD algorithms,
we can gather this sort of data from surveillance cameras. [22] [48] Afterwards, we
can either use them directly using an image-based deep learning method or trans-
form to formulation that can process numerical data. [12]. We discovered a hybrid
approach, that incorporates YOLOv5 with LK optical flow, to more accurately pin-
point the matter. [20], [25], [48]. Together, these two methods offer benefits in
data processing speed and accuracy. It moreover outperforms other object detec-
tion methods such as HOG, SSD, RCC, Fast R-CNN, Faster R-CNN.[21], [30], [35],
[48].

Whenever it comes to anticipating traffic status, [36]AI-based methods [31], [42]

can deliver the most precise prediction to a platform. However, a comparable but
IoT-based system has been deployed and investigated more on the traffic control
system in Bucharest’s congested metropolis.[36]. This study, on the other hand,
will show an updated version of a comparable model that will employ image-based
data rather than IoT-based devices. Algorithms such as the Lucas-Kanade (Lk)
optical approach, for example, can predict speed and anticipate movement of vehi-
cles. [21], [30], [48]. CNN, a deep learning-based model, produces the best results

3
when there is a huge dataset. [46]. It is, in fact, frequently employed in traffic
speed prediction.[39] and traffic flow prediction[34].For these circumstances, a mod-
ified CNN with LSTM is also utilized [35]. This method is applicable among most
urban places, even though in this case, we will just address the traffic issue in Dhaka.

To be more specific, the article covers what we will use and analyze the obtained
data to enable for platform reconfiguration [38] for optimal traffic flow [19]. Accu-
rate forecasts in a complex system like urban traffic flow are a difficult undertak-
ing, especially when random disruptions from the environment are present[12], [26].
Moreover, in this example, meteorological circumstances have a role. [26]. We will
also work past these barriers to develop a strong self-adaptive traffic management
system.

1.1.2 Research Objectives

Automated traffic signal control, with optimized route suggestions and provide traffic
flow, is our main goal for this topic. We aim to apply an image based model with
the accuracy of the robust IoT based model. We set our standards to that level,
although achieving 15% of it is more than enough for our purposes. Likewise, we
also target to achieve this model by implementing a framework as simple as possible
so that the system can be implemented as easily as possible. Such high standards
may not be gained in this paper, but we will follow up with our works to see an end
to finally establish a practical and predictable mode. The key goals of our research
are-

1. To gain knowledge if ML-AI based models optimize traffic flow

2. To get insight into how object algorithm detects density

3. To build a smart ML based flow prediction algorithm

4. To judge the model’s performance based on the precision it provides

5. To compare it with flagship models to assess the practical usability of our

model

4
1.2 Thesis Orientation
The following chapters of the paper maintains the order as described below: Chapter
2 discusses literature review, the architecture we are looking for and similar in the
field and existing methodologies and gives a strong analysis of the topics and related
information that has relevance with our work. Chapter 3 presents the methodology
of our work which includes the workflow, proposed model and its details inner work-
ings. Chapter 4 is where we practically attempt to build and implement the model
and results of our findings, as well as compare them with similar models of the same
criteria. Finally, Chapter 5 contains the concluding notes of the thesis paper so far
and discusses future plans regarding the thesis.

5
Chapter 2

Literature Review

Solving a complex traffic congestion problem like this is not easy. For this, we looked
into already existing researches regarding the topic. It turns out that approach to
solving traffic congestion is not a novel idea at all. To be precise, there has been at
least a thousand researches conducted solely optimizing traffic flow. Optimization
of traffic flow is not an easy feat, and we also don’t intend to do so. What we want
to achieve is a solution that Bangladesh can easily welcome. Automated solutions
are obviously our main goal but if that requires implementation of infeasible and
expensive infrastructure then it’s safe to say that such a research would go in vain
as the country cannot benefit from it too much. Therefore, IoT based solutions are
to be avoided to implement a feasible solution for the country that is image based
solutions.

To get started with our research, we started searching for documents in popular
site like Scopus, Google Scholar. We search for “Traffic Optimization”, “Intelligent
Traffic Control System”, “Smart Traffic Control”, “Image based Traffic Solution”.
We got a lot of researches and shortlisted to about 1200 documents. Now we went
into advanced filter where we searched for “Image Based Solution”. This helped
us narrowed down to 120 documents. Then we looked for traffic optimization and
traffic solution researches, narrowing down to 80 documents. We also had to remove
some of the false positives that included RFID/IoT/V2I infrastructure. With this
filtration, we managed to drop to 60 researches and article that we deemed worthy
of looking into for our research paper.

2.1 Intelligent Traffic System Control (ITSC)

Commonly used for traffic management, it mostly uses managetic detector and sig-
nal systems using loops [2]. These generally have the capability to control traffic

6
dynamically and optimize traffic flow. It intercepts traffic condition some signal
cycles ahead and directs the flow accordingly to avoid gridlocks. This system is an
integral part of smart city concepts [19]. Two important concepts are to be noted
here, SCOOT [2] system in Ontario and SCAT [1], both are ITSC concepts, based
on dynamic traffic coordination [37]. These systems are extremely competent in
getting accurate traffic flow.

2.2 Used Architecture

The feasible approach that we are planning to implement are more than capable to
solve the problem in our opinion. The approaches we will be following are brilliant
in their own respect field but they have never been used together in a single system.
The approaches we will considering are below:

We will use the following 3 layer architecture for our task that we think would more
than appropriate to have a go at providing optimal solution at our problem.

1. Object Detection Algorithm: The YOLO algorithm is one of the most

widely used object detection algorithm [20], [22], [48]. It uses backpropagation,
convolutional layer, CSP and spatial pyramid pooling for object detection.
YOLO was in action more than eight times [4], [6], [27], [47], [48], among our
literature review. So we can safely presume it to be the most formidable model
for object detection.

2. Optical Flow : Speed detection is a crucial part of our system and optical
flow seems the most comfortable choice. Mentioned 5 times in literature review
[18], [19], [45], [12], this model can use frame differencing to calculate global
speed of a group of traffic more specifically, congestion.[48].

3. Numerical Simulations Module : This module is used in coherence with

above two models to indicate the flow stream of traffic and suggest optimal
route. Using fluid dynamics, it approaches to solve the problem in mathemat-
ical order. [32].

7
2.3 Related Works
This section provides a detailed review of previously completed researches where
smart or intelligent traffic models have been implemented. To be precise, we are re-
viewing mostly image based data, but we also review RFID/IoT/V2I infrastructure
where we extracted and tried to bring up relevant information regarding the case
we are solving.

In the past few years, ordinary computers’ computing capabilities have risen dra-
matically, enabling developing and testing numerous RL and DL based algorithms
to handle traffic congestion problems [3], [7]. The first entirely self traffic signal uses
a Q-estimation network to dynamically learn how to control-based on the Cerebellar
Model Articulation Controller (CMAC). [6]. Silva et al. and Oliveira et al. [49] also
presented a context detector in combination with RL. [8].

Nevertheless, in the last 20 years, artificial intelligence and machine learning have
been explored extensively in the domain of traffic light management. [31], [36], [37],
[42], [44]. Mikami et al. introduced a distributed reinforcement learning system
that employs a genetic algorithm, popularly known as the Q-algorithm, to properly
improve a system’s traffic flow. [3]. Later, empirical correlations based on fuzzy
logic and neural networks, such as Bingham’s projected RL for parameter search,
were demonstrated. [4], Choy’s RL-based adaptive large-area cooperative system
[5]. Nevertheless, because processing capacity was insufficient to do such tasks, the
majority of these preliminary works were limited to theoretical fields.

Depending on these studies, a Hybrid Petri Net model [36] was developed that relied
on traffic-dependent phase lighting. Through a control architecture, this was accom-
plished at a reactive level in Bucharest. [28], [34]. In a number of publications, the
topic of smart cities with ITSC integration was also discussed.

Using deep learning approaches [31], [43], the study states [36] an upgraded version
of the presently used ITSC systems that can forecast traffic circumstances 5-10 sig-
nal cycles ahead of time and dynamically change the signal light.

This system is critical to our work since we will be working on producing an opti-
mized version of the recommended architecture with help for overcoming our coun-
try’s technical shortcomings.

8
Chapter 3

Methodology

The focus of this research is to develop an entirely automated traffic management

system. To do our tasks efficiently, we must first obtain photographs from surveil-
lance systems. To execute out this task, we must first collect raw picture data and
label it into various train and test sets, after which we would train our suggested
model using the train set.Likewise, we inject photos into our algorithms and test
them with real-world data to ensure the correctness of the projected traffic flow.
Ultimately, we compare its precision to that of other systems in order to determine
the best one. Figure 1 depicts a high-level overview of the entire process.

This research aims to present a deep learning based AI model that can detect con-
gestion and the motion of it, and can predict its next destination. To achieve this,
first we need raw image data that we can use to build our neural network model. We
will need a properly labeled dataset and, using the train set, we will train the model.
Then using motion detection algorithm we will prepare it for the prediction part. To
easily describe our works, we have divided the whole process into 5 sections. These
are:

1. Dataset Collection: Due to the ongoing pandemic, we couldn’t collect first

hand or primary data. Thus, we had to settle for secondary data collected
from the internet. We will utilize this dataset to train our model and run
further validation.

2. Data Pre-Processing: In this step, we prepared the dataset for our use by
applying some color filter, cropping and other modifications which we assume
will increase the accuracy of tests.

3. Training the Object Detection Model: In this phase, we will train and
test the object detection model, which will be later used for congestion detec-
tion.

9
Figure 3.1: Workflow diagram of the proposed model

10
 4. Adding Flow Detection: Here, we use the optical flow model to detect the
traffic flow, which will help us detect the congestion.

5. Add the optimization module: In this stage, we add a LWR traffic flow
model

6. Run Experiments and Comparisons: In this stage, we run our model in

parallel with other image based AI models and leverage the capabilities of our
model with others.

3.1 Data Collection

For simulating our ITS, we need data for two of our models. Firstly, to train the
object detection model. Secondly, to test the ITS simulation. These two datasets
are discussed briefly below.

3.1.1 Object Detection Data

Since we are working to solve a problem regarding the traffic scenario of Bangladesh,
commonly used datasets of images or traffic couldn’t be used as they were providing
incorrect result most of the time. Thus, we had to look deeper to find dataset col-
lected by our countrymen. We collected a dataset that was uploaded by faculties of
Green University. This dataset contained images of city traffic in various locations
inside the Dhaka division. This particular dataset contained 3503 distinct labeled
images of vehicles in various areas of Dhaka city. This is a labelled dataset that
contains 21 labels which indicates the vehicle type. The images were labeled using
.xml format, which is not suitable for use with YOLOv5[49]. Thus, we had to change
the label files to .txt format using an online tool called Roboflow. Following is a
brief overview of the dataset original dataset:

Table 3.1: Dataset Summary

Number Of Classes Total Images Train Set Test Set

21 3503 3303 200

3.1.2 Traffic Simulation Data

After implementing our proper model, we need some video data to evaluate the
performance of our model. Our goal is to at least reach the accuracy of the IoT

11
models already in existence. For this module, we decided to collect first-hand data
by going to several over bridges of Dhaka city and recording 4 sided footage. Since
over bridges are usually close to an intersection, it represents a more suitable po-
sition of where actual CCTV cameras would be. We opted for video data for this
phase, since we will only be doing inferencing to check the accuracy of our model.
Following is a brief overview of our collected dataset.

Table 3.2: Collected Data For Inferencing

Videos Total Video Size(GB) Compressed Size(GB) FPS

8 10 1.1 24

3.2 Data Preprocessing

Since this is a dataset collected from the internet, a lot of changes had to be made
in order to properly suit with our needs for the proposed system. Also, the dataset
require some standard preprocessing anyway due to a lot of noisy or unusable images.
The following preprocessing tasks were performed to prepare the data for an optimal
condition.

1. Cropping and Resizing: The images were captured with either camera or
smartphone, the framing of which does not match with CCTV format. Since
our goal is applying this model for surveillance systems, we cropped and resized
the image to appropriate format, removing unnecessary parts, only keeping the
road in focus.

2. Desaturating the Images: Most surveillance cameras use monochrome color

format. Especially if it used infrared. Thus, to match the dataset closer to our
expected inferencing data, we desaturated (turned monochrome) those images
that were RGB.

3. Adding Noise: To match CCTV image, we added slight amount of noise

which represent the color and footage type of surveillance systems.

4. Re-Annotating the Images: Since we cropped the images for more focused
and narrow field of view, we had to re-annotate some images.

5. Converting to Text Files: The bounding box information was recorded in

.xml files. To use with YOLOv5, we needed to convert it to .txt file. For

12
 this we used the same online tool, Roboflow, that we also used for our image
manipulation in the above steps.

6. Removing Labels During Annotation: Instead of keeping all 21 labels,

we narrowed it down to only 1. Since our goal is vehicle identification, we just
used only 1 label, “vehicle”, which will increase the F1-Score of the model.
This we further discussed in the Implementation section.

As we can see some preprocessing steps might require us to remove some images
which don’t provide required information for the model and may decrease the ac-
curacy of the model, after these preprocessing steps our dataset now looks like the
following:

Table 3.3: Dataset Summary

Number Of Classes Total Images Train Set Test Set

1 2912 2912 0

Here is the data preprocessing steps shown on a sample image:

Figure 3.2: Preprocessing Steps Example

13
Although we chose the whole dataset for our training, we decided to collect primary
data later on for the cross validation and inferential statistics.

3.3 Model Architecture

Having a labeled dataset in our hand, we can start training the model through
supervised classification. First, we used the benchmark dataset, Coco128, to see
how well the model performs. As expected, the inference wasn’t accurate enough.
Thus, we proceeded with the baseline model and trained it with our dataset. We
trained the model on several parameters to get the maximum precision possible.
Our constructed model are being discussed below:

1. Object Detection Layer: First we need to detect congestion on the road

before proceeding further with motion detection and congestion prediction.
For this, we used YOLOv5 for vehicle detection. Based on the intensity of the

Figure 3.3: Yolov5 Inferencing Steps

detected vehicle, the model decides whether there is congestion on the road or
not.

14
 2. Flow Detection Layer: If there is congestion, the model then proceeds to
detect motion of the vehicle (ultimately the whole congestion). This layer is
built using a modified LK Optical Flow model. This can detect vehicle motion
by frame differencing and interpolation of pixels. The mathematical formula
that is working in LK Optical [23]:

ft,t−k = [u v 0]T (3.1)

Additionally, to determine the swarm success rate, we need the summation

of all the vehicles bodies moving in a single frame. Thus, we can get the
macroscopic flow of the traffic using the following formulation:

N
1 X
SR = (F Mi > TF M ) (3.2)
N i=1

[23]

3. Numerical Simulation Model: For this layer, we will be using the LWR
model. The LWR model used the fluid dynamics approach to describe traffic
flow in a single road[17]. This scalar hyperbolic conservation law can predict
the flow of the traffic. More described below.

3.3.1 Global Speed Detection Algorithm

To detect global speed, we will a use a modified LK Optical Flow algorithm. For
that, we will use an algorithm with YOLOv5. There are 2 reasons for choosing this.
First, the ease of implementation. We do not have to work on another sophisticated
deep learning network. Also, implementing the algorithm is much more convenient
for us.
Given the ROI and YOLOv5 optimally detects all the vehicles to an acceptable per-
centage, we can safely assume with the algorithm above we can detect the speed of
the vehicles. The above algorithm is basically a simplified algorithm of the following
equation [23]:

f (x, y, t = f (x + dx, y + dy, t + dt) (3.3)

15
 Global Speed Detection Algorithm [23]
1. Using LK Optical Flow method, we can differentiate the
frame’s from first frame to the next (from πtoπ + 1). Using the similar function,
we can denote as πr from the position of π + 1.

for i = 1; i ≤ n; i++
do
if d = |π − πr| < Td
then:
retain pi
else:
discard pi
# construct a coordinate, computing θ with respect to π → π + 1
if 90 ≤ 0 ≤ 140,
then:
reserve corner point
else:
discard corner point
done:
# Compute mean value of N frame speed as global vehicle speed V

3.4 Fluid Dynamics Approach

3.4.1 Mathematical Formulation
The LWR model is a widely praised, popular model for traffic flow simulation. It is
based on fluid dynamics approach where it defines the relation between density(k),
flow(q) and local speed, V , of a single road. The formulation is as follows[17]:

 δρ + δq = 0
δt δx
(3.4)
ρ (x, t = 0) = ρ (x)
0

Here, ρ ϵ [0, ρmax ] defines the density at place and x and time t, max indicating
maximum possible density. Moreover, q also represents the vehicles flux (number of
vehicles across a point in a unit time). We can calculate flux using:

q(ρ) = ρv (3.5)

v referring to average speed of vehicles

Assuming vehicle speeds get adjusted depending on surrounding conditions, we can

formulate v as following:

16
  
ρ
v (ρ) = 1 − (3.6)
ρmax
vm ax represents maximum speed where ρ approaches 0
Therefore, the model approaches to define an initial-value problem (Riemann Prob-
lem),

ρ , f orx ≤ x
L 0
(3.7)
ρ , f orx > x
R 0

3.4.2 Computational Functions

Described in Erwin B Setiawan et al., [23] considering the finite volume method for
solving the LWR, they formulated the stability of the numerical solution using the
following, cite49, cite52

δx
δt ≤ minj=1,...,M ρn
(3.8)
|1 − j
ρmax
|vmax ρnj

The probable solutions can be mapped in the following table, the Riemann solution
for the LWR model[17],

Table 3.4: Riemann Solution Mapping

n+1/2
Case Structure Of The Solution ρj+1/2
ρj ≥ ρj & ρnj < ρmax /2
n n
Rarefaction wave heading to the right ρnj

ρnj ≥ ρmax /2 & ρnj+1 ≤ ρmax /2 Rarefaction wave heading to the left ρmax /2

ρnj ≥ ρnj+1 & ρnj+1 ≥ ρmax /2 Rarefaction wave to both direction ρnj+1

ρnj < ρnj+1 & ρnj + ρnj+1 ≤ ρmax Shock wave heading to the right ρnj

ρnj < ρnj+1 & ρnj + ρnj+1 > ρmax Shock wave heading to the left ρnj+1

These 3 modules are key part of our traffic simulation system. Since we solely are
relying on the architecture at hand, we strongly assert that the system we are aiming
for will not only be highly modular but also provide significant results in various
tech stack prepared for real time traffic simulation and intelligent traffic systems[9].

17
Chapter 4

Implementation and Results

So far, we have discussed how the model works theoretically and approaches that
should be taken into account to get optimal results. In this section, we proceed
to implement the models practically to build the actual ITS with self adaptive re
correction capabilities. We will try to some extent to mimic the behavior, however
we won’t approach to completely build the model as it is not possible given the
resources we had at hand during these implementations.

4.1 Implementation Roadmap of Proposed Mod-

els
In this section, we describe how we implemented the proposed ITS to some ex-
tent using popular object detection libraries and other python libraries such as
PyTorch. We have discussed our implementations methods below. image model
training python libraries, PyTorch and Keras. We lay out our roadmap for imple-
menting the model below.

1. Object Detection Layer Construction: For this part of the model, we

used a popular object detection model called YOLOv5 and trained it on the
dataset described in section 3.1.1. We will use online notebooks because of
their high GPU capabilities. We will train our model for a certain epoch till
we get an acceptable accuracy to proceed to our next module[48].

2. Flow Detection Layer Selection: We chose a modified version of the LK

Optical Flow model[21] to detect the congestion. For congestion detection we
need to learn the traffic flow[19],[38], to learn whether the traffic is moving or
not. LK optical flow will be used to detect that traffic flow.

3. Numerical Simulation Of Traffic Flow: To identify the relation between

flow, density, local speed we will use LWR Traffic Flow model.[17] By using the

18
 finite method for the model that we formulated and described in this section
3.4.2, we can an estimated simulation of traffic and implement the prediction
model.[24]

4.2 Object Detection Layer

The first module for our Self Adaptive Intelligent Traffic System is to detect vehicles
on the road, as well as the number of cars. This detection will help us find out the
number of vehicles as well as the density of the vehicles using the ROI approach.
Using the dataset (section 3.1.1, we started training our YOLOv5 model, which is
based on a sophisticated Neural Network design.[25]

4.2.1 Training The Object Detection Model

Our goal for this training was to achieve at least 80% detection rate in detecting
vehicles of Dhaka City streets from grainy/noisy footage. We obtained the following
results on varied epochs. We have laid down our test parameters below, as well as
the achieved. Before starting the test, we chose a pre-trained YOLOv5 large model
which is trained on COCO128, instead of starting from scratch.
Following are the parameters of our Object Detection Model training

train.py --img 640 --batch 16 --epochs 100 --data data.yaml

--weights yolov5l.pt --project vehicle-detection-yolov5 --save-period 10

It took us approximately 6 hours to complete all 100 epochs. We logged our data
using an online tool, and kept track of the results that we were getting.

The summary of our achieved results are described in the table below 4.2:

Table 4.1: Training Set Results

No. Of Epochs mAP Recall Precision Training Loss

20 75.2% 72.02% 86% 2%
50 80.65% 70.80% 85.92% 1.8%
100 82.11% 73.99% 86.73% 1.8%

A graphical overview of the achieved results are shown in following:

19
 Table 4.2: Validation Set Results

No. Of Epochs mAP Recall Precision

20 70.2% 72.02% 83.33%
50 81.65% 76.80% 85.92%
100 82.11% 74.99% 86.73%

Figure 4.1: Training Precision

Figure 4.2: Training Recall

20
 Figure 4.3: Validation Loss

Figure 4.4: Mean Average Precision

21
After the 100th epoch, we saw a sudden and unexpected drop in accuracy for the next
few epochs. But after that the graph stabilized. But we didn’t see any better results
than the 100th epoch provided. After 10 more epochs, we noticed the validation
loss gradient was 0. Meaning, the model wasn’t improving any more and after 10
more epochs we noticed an increase in the validation loss graph, indicating that the
model has started to overfit. Thus, we proceeded with the model that we received
after 100th epoch, which we deem has the best accuracy of 82.11%. The precision
gradient over time can be seen in 4.4.

Of course, providing additional data would most certainly improve the F1-score.
However, we didn’t bother to invest any more of our time for this layer, as greater
than 80% is satisfactory for our purpose.

4.2.2 Inferencing and Results

After training, we set on to collect some new data for inferencing and checking if the
model provides sufficient detection as we need. It turns out that the trained model
works as we planned. We modified the YOLOv5’s detect.py file slightly. We added
the code for ROI and vehicle counting and outputting the values in a file for further
use. An example of inferencing working properly using video footage is shown in
figure 4.5.

Figure 4.5: Inferencing Results of Trained Model

22
4.3 Optical Flow Layer
In combination with YOLOv5, LK Optical flow can provide the global speed of ve-
hicles using frame differentiation, It can be more clearly seen in the figure below:

Figure 4.6: Inferencing Results of Trained Model

Using this algorithm along with YOLOv5, we can run some inferencing on our col-
lected dataset, the results are described in comparison with a previous test in table
4.3 we get the demonstration of our achieved results in the figure 4.6.

Table 4.3: Optical Flow Accuracy

Method Detection Accuracy Detection Time Speed Accuracy

Faster R-CNN 81.25% 233ms 85.5%
YOLOv5 82.38% 184ms 86.3%

Visibly, we can detect a vehicle’s speed from this data and send this data to Numer-
ical Simulations Layer for further actions. Since, the LWR model requires density
and local speed, we have all the necessary data to get an estimated simulation and
gather prediction data.

23
4.4 Numerical Simulations Layer
With the initial problem given Riemann problem, defined in 3.4.2, the convergence
of finite volume method for solving LWR model, we can calculate the norm error
values from the formulation provided by Erwin B. Setiwan et al., where L2 is the
error ratios,
Z l
||ρ(x, t) − ρexact (x, t)||L2 = ( [ρ (x, t) − ρexact (x, t)]2 dx) (4.1)
0

The benefit of calculating the L2 norm values is that we can get an estimation of
errors before hand. By calculating the error we can adjust the offset of our model.
This will widely vary based on the results the previous 2 layers provide.

4.4.1 Test Results

In this first test, we try to simulate the traffic light triggered shock wave. We took
data from our first video (following is a snapshot of the image) where we estimated
the road length to be approximately 150 m. Globally we consider x = 0 m and cars
are coming from 150 m. Since the road is empty, the traffic light is currently at red.
From the data, we found ρm ax = 0.25v/m vm ax = 30km/h. Vehicle density is
calculated by calculating the area from the ROI. Now we have ρL , ρR . Therefore,
we can use the following formulation to get an estimated simulation.[17]

ρ , if (x − x )/t < c
L 0 s
(4.2)
ρ , if (x − x )/t > c
R 0 s

Here. cs = [q(ρL ) − q(ρR )]/(ρL − ρR )

If we run this consecutively for our test dataset, we can see how the traffic signals
work during different traffic load. Based on the achieved ration, the traffic lights
will act accordingly to adjust traffic flow.

 ρ , if (x − x0 )/t < λ(ρL )
 L


1
ρ (1 − vx−x
2 max
0
max t
), if λ(ρL ) < (x − x0 )/t < λ(ρL ) (4.3)


ρ , if (x − x )/t > λ(ρ )

R 0 L

Here, λ(ρ) = (1 − 2ρ/ρmax )vmax

24
We have recorded our test results in table 4.4:
Table 4.4: L2 norm values

M ||ρnum − ρexact ||2L Ratio

20 0.135 -
40 0.100 0.36
80 0.903 0.48
100 0.711 0.56

In the figure 4.7, we have tried to show a simplified version of our recorded results.
We have used the finite volume method against the exact solution. As you can notice
in the figure that it varies significantly when shock waves and rarefaction waves are
registered.

Figure 4.7: Testing Final Volume Method Against Exact Solution

These results are a numeric representation of what we would call congestion and
flow. Of course, it does not seem as intuitive at first glance because it is solely
purposed for machines to operate on this data. Additionally, we will have a closer
look at how the traffic light simulation will work in the next section.

4.5 Traffic Light Simulation

If we were to theoretically explain how the light simulation works, we first assume
a scenario in a single lane road where initial density is ρ0 , approaching the traffic
light, the cars stop. gradually over time, we see an increasing number of cars, and

25
our YOLOv5 model signals that it detects congestion because the vehicle numbers
have passed a certain value. First, let us define the initial variables in table 4.5,

Table 4.5: Simulation Parameters

Variable Parameter
l Length of computation domain
ρl Vehicles flux
vmax Max Speed
vR Vehicle Speed
ρ0 Initial Density
ρmax Max Density
δt Time Step
δx Cell Size

Assume a hypothetical sceniario such as the figue 4.8At ρ0 , we have our traffic signal
giving red. Because our object detection layer is not detecting any congestion. As
time passes, ρ0 → ρmax , our object detector model wakes up the optical flow model
for a scan. Now, if the cars are to flow at a given satisfactory speed, it is safe to
assume that the pile up hasn’t occurred yet. However, given, the optical flow layer
detects low speed, meaning vr = 0 , then the numerical simulations model will be
triggered.

Figure 4.8: Testing Final Volume Method Against Exact Solution

With the trigger of the numerical simulation model, it gathers the real time data
from other traffic around it. If the other lights have been dormant for a while, it
seeks for the latest data. With the data collected, the simulation model can simulate
traffic flow and seek for vR = vmax and ρmax → ρ0 .
We can expand this problem to more than one traffic light without any major com-
plications. These traffic lights can communicate with each other and will try to find

26
out the most optimal route for each case. No, traffic light will show green at the
same time because, it will only react to the lane with maximum flow.

27
Chapter 5

Conclusion

Excessive congestion is not uncommon in countries such as Bangladesh, which has

a massive community and a great amount of automobiles. Road congestion not just
causes financial loss, but it also has a detrimental influence on human health and
wastes a huge amount of resources. Under such conditions, implementing an adap-
tive intelligent traffic system for handling this problem would be a huge ’quality of
life’ benefit for all metropolitan residents. Another important issue is that increased
computer capacity has permitted the creation of a plethora of AI-based models that
can already execute at a significantly quicker rate. Undoubtedly, AI-based deep
learning techniques have made significant progress in tackling such types of complex
challenges. Even with that in mind, this research article outlines our endeavor to
create a self-adaptive ITSC system that can regulate traffic intelligently based on
estimates of traffic conditions in the nearish term. As a consequence, it can control
traffic by delivering the appropriate set of instructions via traffic signals and giv-
ing precedence to crowded lanes. Finally, we compare different AI-based methods
with statistics-based approaches in an attempt to discover the best model for our
proposed model.

5.1 Future Works

In the future, we plan to extend our work by implementing a centralized AI model,
preferably based on Deep learning network that can take instantaneous decision
based on the data that the traffic lights are providing which are based on numerical
simulations. We have looked at such models and also found some interest in Google’s
graph neural network technology suggested by our faculty. Such a robust system
can be applicable anywhere, in any infrastructure, and can later be swapped with
smart city features if required. We plan to extend our dataset and realm of testing
to make our already established plan more robust, as well as centralized AI hive
mind technology to predict and optimize traffic flow in the whole city.

28
Bibliography

[1] P. Hunt, D. Robertson, R. Bretherton, and M. C. Royle, “The scoot on-line

traffic signal optimisation technique,” Traffic Engineering & Control, vol. 23,
no. 4, 1982.
[2] P. Lowrie, “Scats, sydney co-ordinated adaptive traffic system: A traffic re-
sponsive method of controlling urban traffic,” 1990.
[3] S. Mikami and Y. Kakazu, “Genetic reinforcement learning for cooperative
traffic signal control,” in Proceedings of the First IEEE Conference on Evolu-
tionary Computation. IEEE World Congress on Computational Intelligence,
IEEE, 1994, pp. 223–228.
[4] E. Bingham, “Reinforcement learning in neurofuzzy traffic signal control,”
European Journal of Operational Research, vol. 131, no. 2, pp. 232–241, 2001.
[5] M. C. Choy, D. Srinivasan, and R. L. Cheu, “Hybrid cooperative agents
with online reinforcement learning for traffic control,” in 2002 IEEE World
Congress on Computational Intelligence. 2002 IEEE International Conference
on Fuzzy Systems. FUZZ-IEEE’02. Proceedings (Cat. No. 02CH37291), IEEE,
vol. 2, 2002, pp. 1015–1020.
[6] B. Abdulhai, R. Pringle, and G. J. Karakoulas, “Reinforcement learning for
true adaptive traffic signal control,” Journal of Transportation Engineering,
vol. 129, no. 3, pp. 278–285, 2003.
[7] D. de Oliveira, A. L. Bazzan, B. C. da Silva, et al., “Reinforcement learning
based control of traffic lights in non-stationary environments: A case study in
a microscopic simulator.,” in EUMAS, 2006.
[8] B. Silva, D. d. Oliveira, A. Bazzan, and E. Basso, “Adaptive traffic control
with reinforcement learning,” in Proceedings of the 4th Workshop on Agents
in Traffic and Transportation (AAMAS 2006)(May 2006), ALC Bazzan, BC
Draa, and F. Kl, Eds, 2006, pp. 80–86.
[9] J. J. S. Medina, M. J. G. Moreno, and E. R. Royo, Evolutionary computation
applied to urban traffic optimization. IntechOpen, 2008.

29
[10] L. Xiao and Z. Wang, “Internet of things: A new application for intelligent
traffic monitoring system,” Journal of networks, vol. 6, no. 6, p. 887, 2011.
[11] L. Hernández, C. Baladrón, J. M. Aguiar, et al., “A study of the relation-
ship between weather variables and electric power demand inside a smart
grid/smart world framework,” Sensors, vol. 12, no. 9, pp. 11 571–11 591, 2012.
[12] D. Stefanoiu, J. Culita, and F. Tudor, Experimental approaches in process and
phenomena identification, 2012.
[13] M. S. Ali, M. Adnan, S. M. Noman, and S. F. A. Baqueri, “Estimation of
traffic congestion cost-a case study of a major arterial in karachi,” Procedia
Engineering, vol. 77, pp. 37–44, 2014.
[14] A. Nkoro and Y. A. Vershinin, “Current and future trends in applications of
intelligent transport systems on cars and infrastructure,” in 17th International
IEEE Conference on Intelligent Transportation Systems (ITSC), IEEE, 2014,
pp. 514–519.
[15] P. S. Paul et al., “Traffic density estimation and flow control for video surveil-
lance system,” Ph.D. dissertation, BRAC University, 2014.
[16] Y. Yuan, H. Van Lint, F. Van Wageningen-Kessels, and S. Hoogendoorn,
“Network-wide traffic state estimation using loop detector and floating car
data,” Journal of Intelligent Transportation Systems, vol. 18, no. 1, pp. 41–50,
2014.
[17] E. B. Setiawan, D. Tarwidi, and R. F. Umbara, “Numerical simulation of
traffic flow via fluid dynamics approach,” International Journal of Computing
and Optimization, no. 3, pp. 93–104, 2015.
[18] Y. Ding and X. Fu, “Kernel-based fuzzy c-means clustering algorithm based
on genetic algorithm,” Neurocomputing, vol. 188, pp. 233–238, 2016.
[19] K. Nellore and G. P. Hancke, “A survey on urban traffic management system
using wireless sensor networks,” Sensors, vol. 16, no. 2, p. 157, 2016.
[20] J. Redmon, S. Divvala, R. Girshick, and A. Farhadi, “You only look once:
Unified, real-time object detection,” in Proceedings of the IEEE conference on
computer vision and pattern recognition, 2016, pp. 779–788.
[21] A. Ranjan and M. J. Black, “Optical flow estimation using a spatial pyra-
mid network,” in Proceedings of the IEEE conference on computer vision and
pattern recognition, 2017, pp. 4161–4170.
[22] J. Sochor, R. Juránek, and A. Herout, “Traffic surveillance camera calibration
by 3d model bounding box alignment for accurate vehicle speed measurement,”
Computer Vision and Image Understanding, vol. 161, pp. 87–98, 2017.

30
[23] J. Huang, W. Zou, J. Zhu, and Z. Zhu, “Optical flow based real-time moving
object detection in unconstrained scenes,” arXiv preprint arXiv:1807.04890,
2018.
[24] ——, “Optical flow based real-time moving object detection in unconstrained
scenes,” arXiv preprint arXiv:1807.04890, 2018.
[25] J. Redmon and A. Farhadi, “Yolov3: An incremental improvement,” arXiv
preprint arXiv:1804.02767, 2018.
[26] B. Deb, S. R. Khan, K. T. Hasan, A. H. Khan, and M. A. Alam, “Travel time
prediction using machine learning and weather impact on traffic conditions,”
in 2019 IEEE 5th International Conference for Convergence in Technology
(I2CT), IEEE, 2019, pp. 1–8.
[27] E. Ferrara, L. Fragale, G. Fortino, et al., “An ai approach to collecting and
analyzing human interactions with urban environments,” IEEE Access, vol. 7,
pp. 141 476–141 486, 2019.
[28] W. Ghazali, C. Zulkifli, and Z. Ponrahono, “The effect of traffic congestion
on quality of community life,” in 4th International Conference on Rebuilding
Place. The European Proceedings of Multidisciplinary Sciences, 2019, pp. 759–
766.
[29] W. H. Money and S. Cohen, “Leveraging ai and sensor fabrics to evolve smart
city solution designs,” in Companion Proceedings of The 2019 World Wide
Web Conference, 2019, pp. 117–122.
[30] D. Impedovo, F. Balducci, V. Dentamaro, and G. Pirlo, “Vehicular traffic
congestion classification by visual features and deep learning approaches: A
comparison,” Sensors, vol. 19, no. 23, p. 5213, 2019.
[31] A. Miglani and N. Kumar, “Deep learning models for traffic flow prediction
in autonomous vehicles: A review, solutions, and challenges,” Vehicular Com-
munications, vol. 20, p. 100 184, 2019.
[32] ——, “Deep learning models for traffic flow prediction in autonomous vehi-
cles: A review, solutions, and challenges,” Vehicular Communications, vol. 20,
p. 100 184, 2019.
[33] S. Ramteke and B. Gite, “Ai based traffic signal control system,” Traffic, vol. 6,
no. 11, 2019.
[34] W. Zhang, Y. Yu, Y. Qi, F. Shu, and Y. Wang, “Short-term traffic flow pre-
diction based on spatio-temporal analysis and cnn deep learning,” Transport-
metrica A: Transport Science, vol. 15, no. 2, pp. 1688–1711, 2019.

31
[35] T. Bogaerts, A. D. Masegosa, J. S. Angarita-Zapata, E. Onieva, and P. Hellinckx,
“A graph cnn-lstm neural network for short and long-term traffic forecasting
based on trajectory data,” Transportation Research Part C: Emerging Tech-
nologies, vol. 112, pp. 62–77, 2020.
[36] J. Culita, S. I. Caramihai, I. Dumitrache, M. A. Moisescu, and I. S. Sacala,
“An hybrid approach for urban traffic prediction and control in smart cities,”
Sensors, vol. 20, no. 24, p. 7209, 2020.
[37] K. S. Desikan, V. J. Kotagi, and C. S. R. Murthy, “Topology control in
fog computing enabled iot networks for smart cities,” Computer Networks,
vol. 176, p. 107 270, 2020.
[38] A. Essien, I. Petrounias, P. Sampaio, and S. Sampaio, “A deep-learning model
for urban traffic flow prediction with traffic events mined from twitter,” World
Wide Web, pp. 1–24, 2020.
[39] R. Ke, W. Li, Z. Cui, and Y. Wang, “Two-stream multi-channel convolutional
neural network for multi-lane traffic speed prediction considering traffic volume
impact,” Transportation Research Record, vol. 2674, no. 4, pp. 459–470, 2020.
[40] J. J. Vázquez, J. Arjona, M. Linares, and J. Casanovas-Garcia, “A comparison
of deep learning methods for urban traffic forecasting using floating car data,”
Transportation Research Procedia, vol. 47, pp. 195–202, 2020.
[41] J. Wang, J. Li, S. Yan, et al., “A novel underwater acoustic signal denoising
algorithm for gaussian/non-gaussian impulsive noise,” IEEE Transactions on
Vehicular Technology, vol. 70, no. 1, pp. 429–445, 2020.
[42] P. Xie, T. Li, J. Liu, S. Du, X. Yang, and J. Zhang, “Urban flow prediction from
spatiotemporal data using machine learning: A survey,” Information Fusion,
vol. 59, pp. 1–12, 2020.
[43] C. Zhang, “Design and application of fog computing and internet of things ser-
vice platform for smart city,” Future Generation Computer Systems, vol. 112,
pp. 630–640, 2020.
[44] R. Zhang, A. Ishikawa, W. Wang, B. Striner, and O. K. Tonguz, “Using re-
inforcement learning with partial vehicle detection for intelligent traffic signal
control,” IEEE Transactions on Intelligent Transportation Systems, 2020.
[45] A. H. Ahmed and L. Z. Fragonara, “Adaptive intelligent traffic control systems
for improving traffic quality and congestion in smart cities,” 2021.
[46] M. Akhtar and S. Moridpour, “A review of traffic congestion prediction using
artificial intelligence,” Journal of Advanced Transportation, vol. 2021, 2021.

32
[47] S. Majumdar, M. M. Subhani, B. Roullier, A. Anjum, and R. Zhu, “Con-
gestion prediction for smart sustainable cities using iot and machine learning
approaches,” Sustainable Cities and Society, vol. 64, p. 102 500, 2021.
[48] X. Yang, F. Wang, Z. Bai, F. Xun, Y. Zhang, and X. Zhao, “Deep learning-
based congestion detection at urban intersections,” Sensors, vol. 21, no. 6,
p. 2052, 2021.
[49] F. Zhou, H. Zhao, and Z. Nie, “Safety helmet detection based on yolov5,” in
2021 IEEE International Conference on Power Electronics, Computer Appli-
cations (ICPECA), IEEE, 2021, pp. 6–11.

33